Obstacle Recognition for Agricultural UAVs Based on Improved YOLOv8 and Sensor Data Collaboration

To enhance the accuracy and real-time performance of obstacle recognition for plant protection UAVs in complex farmland environments, this paper proposes an obstacle detection method that integrates multi-sensor collaborative perception with a lightweight improved YOLOv8 model. By constructing a triple-perception system and incorporating spatiotemporal registration and dynamic compensation algorithms—along with an attention-convolution hybrid structure, a rotation-sensitive detection head, and a multimodal confidence fusion module—the proposed method significantly improves robustness and detection accuracy under challenging lighting and occlusion conditions. On a self-constructed dataset, the proposed DF-R-YOLOv8 model outperforms mainstream YOLO series models and achieves an inference frame rate that meets real-time control requirements. Experimental results demonstrate the effectiveness and practicality of the proposed method for farmland obstacle detection tasks.